Antonio Duarte

A big one: a turn can now keep going when its model gives out, you can roll back what a turn did to your files, and bots can build up skills of their own.

• Fallback models. When a model hits its usage limit, loses its login, is unavailable or runs out of context, the turn can continue on the next model in your list, API-key providers included. Set it up in Settings Models and agents Fallback models. It's off until you add a model, and you can pin a conversation to its own model.

• Undo a turn's file changes. When a turn changes files, its answer shows the diffs, and "Undo changes of this turn" puts them back, whether the agent used its file tools or ran commands. You review each file first, can undo just one, and can redo. It works with or without Git.

• Bots keep their own skills. A bot can save a procedure it learned as a skill and improve it over time. In the bot's Settings Skills you see each skill's history, restore a version, edit or lock it, or require your approval before a change takes effect.

• One account rotation per provider. Claude and Codex accounts now share a single ordered rotation in Settings Accounts. An account you pick directly stays fixed and stops when it runs out instead of switching. The failover routes you already had become the new rotation.

• Long conversations hold up better. Agents search past messages by what they're about, not only by exact words, and compaction keeps decisions, pending requests and exact identifiers in fixed sections, with the latest part of the conversation kept word for word.

• Sturdier with local models. Ollama, LM Studio, llama.cpp and vLLM report the context window they actually run with, and malformed tool calls from any model are repaired before they fail.

Release notes: https://github.com/antonioducs/m...

This release is about running Maestrly on several Claude or Codex accounts without switching between them by hand.

• Active account in any model selector. Pick Claude Active account or Codex Active account and Maestrly runs on whichever of your accounts is active. You change it in one place: Settings Accounts Active account.

• Use each account until it runs out. Maestrly stays on an account until its quota ends, then moves to the next one and keeps using it, so the reset windows start one after another instead of overlapping. Or choose Return to the active account to go back to it as soon as it resets.

• Everything follows it. Chats, subagents, compaction and image interpretation all use the active account. Picking a specific account still works as before.

• Claude plan renames no longer break your sessions. Conversations and usage limits keep working when Anthropic renames a plan, including Max to Ultra. Existing Claude sessions are rebuilt once from the conversation history the next time you use them, and the history is kept.

• Long conversations switched to Google AI keep your whole message. Maestrly summarizes an oversized history before sending it instead of silently cutting off the end of the prompt. The first response after the switch can take a bit longer.

• Interrupted responses stop cleanly. If Maestrly closes or restarts mid-response, tools and subagent cards no longer spin forever, and the response says it was interrupted.

Release notes: https://github.com/antonioducs/m...

Bots can now hand their heavy work to your other computers, and wait for each other's answers without dropping what they were doing.

• Runners. Turn one of your computers into a runner and your bots can send it heavy jobs like video renders, model inference, large builds or test suites, instead of using the CPU and memory they share with their browsers. The outputs land back in the bot's folder and the conversation picks up with the result.

• Watch and cancel jobs. You see each job's CPU, memory and live log, and can cancel it from any paired computer. Jobs run natively on macOS and Linux, or in a Docker container on macOS, Linux and Windows.

• Easy to set up. On the computer, turn on Run jobs for bots in Settings Runner, then allow the bot in its Settings Runners. A linked bot can run any command on that computer, so only link bots you trust.

• Bots that wait for each other. When a bot asks another bot something, it can now wait for the answer in the same turn and finish the task, instead of reading the reply only on its next turn.

• Steadier long sessions. An approved plan survives context compaction, and Claude no longer claims it can't see an image you just attached after its session is rebuilt.

Runners need the computer, the bot server and the bot environment on v0.30.0, and the mobile app doesn't show them yet. If you run your own bot server, back up the gateway volume before updating it.

Release notes: https://github.com/antonioducs/m...